Abstract
Hallucination is a known issue for neural abstractive summarization models. Recent work suggests that the degree of hallucination may depend on errors in the training data. In this work, we propose a new method called Contrastive Parameter Ensembling (CaPE) to use training data more effectively, utilizing variations in noise in training samples to reduce hallucination. We first select clean and noisy subsets from the training data using different automatic factual metrics. Then, we fine-tune a base summarization model, which is trained on all training samples, on the clean (noisy) subset to obtain an expert (anti-expert) model. Finally, we adjust the parameters of base model by the difference between parameters of the expert and anti-expert models, steering the base model towards the expert model and away from the anti-expert model. Experimental results show that CaPE improves performance across different automatic factual metrics and human evaluation, with the maximum improvement of 16.69% and 15.78% on summary-level dependency-arc entailment accuracy for the XSUM and CNN/DM datasets. The improvement in factual performance does not degrade the performance on other metrics of informativeness such as ROUGE.
Introduction
Neural abstractive summarization models are notorious for "hallucinating" — generating information that sounds plausible but is not supported by the source text. While these models excel at mimicking human language, they often inherit undesirable behaviors from their training data. Standard datasets like XSUM and CNN/DM contain reference summaries written by humans that frequently include "extrinsic" information (facts not present in the original article). When a model is trained on this "noisy" data, it learns that hallucinating is an acceptable part of the task.
Simply filtering out noisy data is not a perfect solution; it significantly shrinks the training set, often leading to "boring" models that lose diversity and fail to recall important information. This research introduces a method to utilize the entire dataset while specifically neutralizing the parameters responsible for factual errors.
Methods: Contrastive Parameter Ensembling (CaPE)
The proposed method, Contrastive Parameter Ensembling (CaPE), also referred to as a Mixture of Factual Experts (MoFE), operates in three main stages:
- Data Selection: Automatic metrics, such as Entity Overlap and Dependency Arc Entailment (DAE), are used to split the training data into "Clean" and "Noisy" subsets.
- Expert Training: A Base Model is first trained on the full dataset. It is then fine-tuned on the Clean subset to create an Expert and on the Noisy subset to create an Anti-Expert.
- Contrastive Ensembling: The models are combined in parameter space to avoid inference overhead. The final model weights ($\theta_{CaPE}$) are calculated as: $$\theta_{CaPE} = \theta_{Base} + \alpha(\theta_{Expert} - \theta_{Anti-Expert})$$
By adding the difference between the Expert and Anti-Expert parameters to the base model, the system is steered toward factual behavior and away from the patterns that lead to hallucinations. This approach generalizes WiSE-FT by adding the "Anti-Expert" component, which proves critical for error reduction.
Experimental Results
The study evaluated CaPE on two benchmarks: XSUM (highly abstractive and noisy) and CNN/DailyMail (more extractive and cleaner).
Factuality and Informativeness
CaPE demonstrated significant improvements in factual consistency across multiple metrics. As shown in the results table, the method achieved a boost in summary-level dependency-arc entailment accuracy of 16.69% on XSUM and 15.78% on CNN/DM.
Notably, unlike simple data filtering which often causes a drop in ROUGE scores (informativeness), CaPE maintains stable ROUGE performance. It effectively corrects factual errors without sacrificing the content density of the summaries.
Human Evaluation and Efficiency
Human annotators confirmed these findings. In head-to-head evaluations on XSUM, CaPE improved factual consistency in 19% of summaries while degrading it in only 14%. Furthermore, because ensembling occurs at the parameter level before deployment, the final model maintains the same inference speed and size as the original base model (zero overhead).
Conclusion
Contrastive Parameter Ensembling (CaPE) offers a robust framework for mitigating hallucinations in abstractive summarization. By leveraging an "Anti-Expert" to subtract noisy parameters, the model achieves superior factual accuracy without the typical trade-offs in informativeness or computational efficiency. Future work may explore applying this contrastive steering to other model behaviors, such as reducing bias or toxicity.
Summary updated